Step 3: Scenes and References
Attach the reference images and clips your model matches for subject, scene, motion and style.
What is the Scenes and References step?
Step 3 is where you attach the visual reference material the model works from — the subject, the scene, the motion and the style it should match. It stays optional: skip it and the model composes a fitting setting from your script and your layout alone.
The step has two versions, and which one you get is decided by the model you locked in at Step 1. Confirm which one you are looking at before you read on.
Which version do you see? Two signals, in order. The section heading in
the editor: Scenes and References is the reference-pool version, while
Set the Scene is the version every other model gets. The tab set:
References plus Generate with AI is the pool version; Image, Video,
Generate with AI, Preset and Color is the other one. The cause is the
model locked in at Step 1 — models that accept a pool of reference material
show the first, and every other model shows the second.
Reference-pool models
These models accept a pool of reference material. Characters, products and scene stills all draw on one shared set of slots rather than each having a private allowance, and the picker shows you how much of the pool is left as you attach things.
| Model | Reference images | Reference clips |
|---|---|---|
| MiniMax Hailuo H3 | up to 9 | up to 3 |
| Seedance 2.5 | up to 30 | up to 10 |
| Seedance 2 | up to 9 | up to 3 |
| Seedance 2 Fast | up to 9 | up to 3 |
| Seedance 2 Mini | up to 9 | up to 3 |
| Wan 2.7 | up to 5 | — |
The model picker at Step 1 lists MiniMax Hailuo H3 under its short name, MiniMax H3.
Whether an attached scene image costs you a slot depends on the model. On Seedance and on MiniMax Hailuo H3 it does — the scene image is folded into the same pool as your characters and products, so it takes one of the slots in the table above. On Wan 2.7 it does not: Wan folds the scene into the text prompt instead and leaves the whole pool free for characters and products.
Why your Step 2 photos appear here
The character and product images you picked in Step 2 draw on this same pool, so they show up in the references strip as read-only tiles labelled "From Character & Product — manage in step 2". They are shown rather than hidden on purpose: the slot counter above the strip counts them, so it always reflects how many slots are genuinely left. To change one, go back to Step 2 — they cannot be edited from here.
Generate with AI
Describe the background you want and the AI generates a reference image from your description. Generating one costs 3 credits, and the result is saved to your brand's asset library so you can reuse it on later videos.
In Easy mode, the × on the References tile clears every reference image at once, not just the last one. The toast that appears offers an Undo action — use it straight away if that was not what you meant.
The list above is today's line-up, not a fixed rule. Any model that advertises a reference-image limit uses this pool — the Step 1 model picker is the source of truth for the model you are actually on.
Every other model
Every model outside that pool keeps the Set the Scene controls, where you
pick a single backdrop from five tabs. Nothing about how these work has changed.
Image — upload your own backdrop: a real location, a branded set, or any image you already have.
Video — attach a clip for the model to match for motion and style. This tab only appears on models that accept a reference clip.
Generate with AI — describe the background you want and the AI generates it. Generating one costs 3 credits, and the result is saved to your brand's asset library for reuse.
Preset — a pre-designed setting: Office, Studio, Park, City, Home, Cafe, Industrial, Minimal — or None.
Color — a clean solid colour or gradient, which suits product-first or text-driven videos.
Setting the scene
Pick a source
Choose Image, Generate with AI, Preset, or Color from the tabs — plus Video, where the model supports a reference clip.
Browse the presets
On the Preset tab, click a setting (Office, Studio, Park, City, Home, Cafe, Industrial, Minimal) or None.
Or describe your own
On the Generate with AI tab, describe the backdrop you want. You need a prompt here before the video can be generated.
A backdrop image is only sent to models that accept one. On a model without image support your choice is folded into the text prompt instead.